Do Landlords Discriminate in the Rental Housing Market...

Do Landlords Discriminate in the Rental Housing Market? Evidence from an Internet Field Experiment in U.S. Cities

Andrew Hanson Department of Economics, Georgia State University

P.O. Box 3992 Atlanta, GA 30302 [email protected]

Zackary Hawley Department of Economics, Georgia State University

P.O. Box 3992 Atlanta, GA 30302

[email protected]

This paper tests for racial discrimination in the rental housing market using matched pair audits conducted via e-mail for rental units advertised on-line. We reveal home-seekers’ race to landlords by sending e-mails from names with a high likelihood of association with either whites or African Americans. Generally, discrimination occurs against African American names; however, when the content of the e-mail messages insinuates home-seekers with high social class, discrimination is non-existent. Racial discrimination is more severe in neighborhoods that are near “tipping points” in racial composition, and for units that are part of a larger building.

JEL: J15, C93

We thank Josh Carter, Jason Cook, Michael Santas, and Dongkyu Won for excellent research assistance. We would also like to thank Spencer Banzhaf, Bentley Coffey, Jim Cox, Paul Ferraro, Dan McMillen, Jon Rork, Kurt Schnier, Geoffrey Turnbull, John Yinger, and Bo Zhao for helpful discussion and comments.

1

I. Introduction

Inequality in housing market outcomes between African Americans and whites is

staggering. African Americans have worse outcomes than whites in terms of housing unit

quality and quality of neighborhood.1 Data from the 2007 American Housing Survey shows that

African Americans are 2 times more likely than whites to have recently seen a rat in their unit;

30 percent more likely to report that the water in their unit is unsafe for drinking and cooking; 60

percent more likely to report a serious crime occurring in their neighborhood in the previous

year; and 2 times as likely to report being dissatisfied with the neighborhood elementary school.

Unequal outcomes between African Americans and whites could be the result of

correlation between race and income, sorting based on the level of local public goods, or

difference in preferences across racial groups. A more sinister (and illegal) source of racial

inequality in the housing market is discrimination against African Americans, the focus of this

paper. This paper identifies discrimination in the rental housing market using matched pair

audits, by contacting landlords via e-mail about rental units advertised through a popular on-line

venue. We highlight the race of home-seekers to landlords through the name attached to each e-

mail inquiry, using names with a high likelihood of association with either whites or African

Americans. We also test how the interaction between race and social class effects landlord

response to e-mail inquiries by altering the type, in terms of the writing style, spelling, grammar,

salutation, and valediction, of e-mail sent to landlords.

1 Unit and neighborhood quality are direct outcomes in the housing market. Indirectly, the consequences of racial segregation in the housing market are also important. See Cutler and Glaeser (1997) for an examination of the effects of racial segregation in the housing market on employment and education outcomes. See Cutler, Glaeser, and Vigdor (1999) for an examination of the causes of racial segregation and how it has changed throughout American history.

2

This paper makes three contributions to the existing literature on racial discrimination in

the housing market. First, we use an on-line venue to conduct an audit-style experiment via e-

mail correspondence with landlords in the United States. E-mail correspondence is

advantageous in an audit-style study because it does not rely on actors, who may have different

appearances, styles, or bring personal bias to the study. Second, by manipulating the language in

e-mail inquiries, we examine the interaction between race and social class. We create two

classes of e-mails, “high” and “low” based on the content of the message and augment our

experiment between races to include between class within race and between class across race

groups. Third, we test for discrimination across neighborhood and housing unit characteristics,

including racial composition.

Overall, our results reveal a net level of discrimination of 4.5 percentage points against

African American sounding names, statistically significant at the one percent level and consistent

with previous studies of racial discrimination in the housing market. When e-mail inquiries

imply the African American is of higher social class, racial discrimination is small and not

statistically different than zero- a unique finding in the literature. When e-mail inquiries imply

that both races are of lower social class we find a larger (6 percentage points) level of net

discrimination against African Americans. The presence and severity of discrimination also

varies across cities in our sample and by neighborhood and unit characteristics. Discrimination

is more severe in neighborhoods that are close to “tipping points” in racial composition as

described in Card, Mas, and Rothstein (2008), and for units advertised as part of a larger

apartment building.

The next section of the paper is a discussion of the previous research on discrimination in

the housing market and places our work in context. Section III describes our experiment.

3

Section IV presents some descriptive statistics of the housing units in our sample and their

surrounding neighborhoods. Section V presents the results of our experiment. Section VI

discusses the robustness and external validity of our results. The final section of the paper

concludes.

II. Previous Research on Racial Discrimination in the Housing Market

The primary method used to test for discrimination in the housing market is an audit, or

matched pair study.2 In an audit study, two subjects (one from the majority racial group and one

from the minority) are matched based on observable characteristics (excluding race) and trained

how to act toward a real estate agent or landlord. The subjects are sent (in random order) to a

landlord or real estate agent’s office to inquire about an advertised housing unit.3 Typically,

subjects will report if they are shown the advertised unit, if they are shown similar units, how

many additional units they are shown, and potentially several other objective measures of

treatment.4

Studies of discrimination using in-person audits include; Yinger (1986), Page (1995),

Ondrich, Stricker, and Yinger (1998), Ondrich, Ross, and Yinger (2000), and Ondrich, Ross, and

Yinger (2003), Zhao (2005), and Zhao, Ondrich, and Yinger (2006). Yinger (1986) examines

the Boston housing market using unique data, the other studies use data from the Housing

Discrimination Study (HDS) conducted by the Department of Housing and Urban Development.

2 There are also several studies that attempt to identify housing market discrimination using price differentials (either from transactions or reported values) between racial groups controlling for observable differences in unit and owner characteristics. Unobservable variables at the owner, unit, or neighborhood level have the potential to confound this method of identification. See Knowles-Myers (2004) for a recent example of this method and for a review of previous studies that identify discrimination using price (or reported value) differentials. 3 Typically, researchers randomly draw housing units from local newspaper advertisements. Geographic coverage of audits has varied substantially, from using a single metropolitan area up to 25 different areas in the same study. 4 See Ross and Turner (2005) for a listing of all measures used in the 2000 Housing Discrimination study conducted by HUD.

4

These studies all find significant discrimination against African Americans. These studies show

discrimination occurs in terms of being told a unit is available, the number of housing units

shown, realtor follow-up communication, and effort on the part of the real estate agent.

Using audits to study discrimination in the housing market has several advantages over

methods that do not have a rigorous control-treatment design. First, because the level of

observation is the landlord or real estate agent, any personal characteristics that may affect

discriminatory outcomes are held constant. Second, audits allow for a direct test of

discrimination in the housing market that is not confounded by discrimination in other markets.

For instance, discrimination in the lending market confounds using sales price differences to

measure discrimination in the housing market. Finally, if done correctly, the race of each auditor

is the only characteristic that varies between members of an audit pair.

Despite the benefits of audits, there are problems with using them to study discrimination

in the housing market (see Heckman (1998) for a detailed description of the problems with in-

person audit studies). Heckman and Siegelman (1993) note that in-person audits rely on how

comparable the actors in an audit are. In order for the audit to be truly unbiased, actors must be

identical along all dimensions except race. Any matched-pair audit using human subjects

certainly violates this assumption, but proper choice and training of actors diminishes the

severity of the problem.

In addition, actors in an audit may bias the study with their own personal beliefs about

discrimination. For example, if actors from one race have prior beliefs about discrimination,

they may be more likely to report discriminatory behavior, or they may act to prompt

discriminatory behavior in subjects. In-person audits are also complicated by the time elapse

5

between visits to the landlord or real estate agent and actors that are not exposed to the same

agent despite visiting the same office and making the same inquiry. These problems are almost

entirely a function of using actors to perform audits in an in-person setting, rather than the audit

design itself.

Ahmed and Hammarstedt (2008) apply the audit technique to housing market interactions

that take place via on-line advertisements and e-mail correspondence. Specifically, Ahmed and

Hammarstedt examine a Swedish housing advertisement website, Blocket.se, to study racial

discrimination between native Swedes and the Muslim minority, and find significant

discrimination toward Muslims. More recently, Ahmed, Andersson, and Hamarstedt (2010) and

Bosch, Carnero, and Farre (2010) study how the interaction between positive information and

race affects landlord discrimination. Ahmed, Andersson, and Hamarstedt (2010) find that while

information (including marital status, employment information, age, and education level) does

increase the response to minority applicants, it does not decrease the difference in response

between native Swedes and the Muslim minority. Bosch, Carnero, and Farre (2010) find

discrimination against the Moroccan minority in Spain, and that positive information increases

the chance of being contacted, but does not eliminate discrimination.

There are two other studies we are aware of that use on-line housing market interactions

in the United States to study discrimination- Carpusor and Loges (2006), and Ewens, Tomlin,

and Wang (2009)- although neither uses an audit style design so they cannot completely control

for landlord characteristics or determine how often landlords treat auditors equally. Carpusor

and Loges (2006) find discrimination against both African American and Muslim sounding

names in the Los Angeles rental market. Ewens, Tomlin, and Wang (2009) study a broader

range of cities than Carpusor and Loges and vary the information supplied to landlords

6

(information about occupation and smoking preference, for instance). They find that African

American home-seekers receive 9 responses for every 10 a white home-seeker receives, and that

including positive information does not affect the response rate difference between races.

Our paper adds to the existing literature on racial discrimination in the housing market in

several ways. First, we apply the audit-style design to an on-line market in the United States.

While previous research has applied the audit-style design to on-line markets in other countries,

or used an on-line venue to study discrimination in the United States, we are the first to combine

these two features. The audit technique is an improvement over other on-line studies of the U.S.

markets as it removes any landlord specific effects and allows us to directly observe landlords

practicing equal treatment or discrimination.

Second, we introduce the notion of social class into our experiments by changing the

language in our e-mail correspondence with landlords. This allows us to test if the interaction

between race and social class is important and if discrimination varies with social class.

Although, we cannot prevent landlords from inferring traits besides race about the names in our

study, we attempt to influence them into inferring something about the social class of the

auditors and test how this matters for response. We also perform several robustness checks,

excluding names by religious affiliation, uniqueness, or differential responses by geography, to

test whether confounding factors associated with the names in our study affect our primary

results. Lastly, we test for differences in discrimination across various unit and neighborhood

characteristics including racial composition, which has not been done with on-line audits.

For a more complete review of the existing literature on racial discrimination in the

housing market see Yinger (1998) and Ross and Turner (2005). Also, see Yinger (1998) and

7

Antecol and Cobb-Clark (2008) for references to studies of racial discrimination in other

markets. Bertrand and Mullainathan (2004) is an excellent example of an audit-style study of

discrimination in the labor market, for a review and references to studies of labor market

discrimination see Altonji and Blank (1999).

III. Experiment Design

Following the audit-style design used by Yinger (1986), Bertrand and Mullainathan

(2004), Ahmed and Hammarstedt (2008), and the studies using the HDS audit data, we design an

experiment to test for differential treatment between African American and white home-seekers

in the rental housing market. In our experiment, we conduct audits through e-mail

correspondence at the landlord level- each prospective landlord receives an e-mail inquiry from

two home-seekers.

We manipulate the racial group (African American or white) of home-seekers through the

name associated with each e-mail inquiry. The sample of names used to manipulate race comes

from Bertrand and Mullainathan (2004), who use Massachusetts birth certificate data from 1974

to 1979 to identify names strongly associated with either white or African American babies to

study discrimination in the labor market. They use the likelihood a baby with a given name is

white, L(W), and the likelihood a baby with a given name is African American, L(AA), to create

a the relative likelihood that a name is white, L(W)/L(AA) or African American, L(AA)/L(W)

The first names we use to represent white home-seekers all have an infinite relative

likelihood measure. The relative likelihood measure is infinite because the Massachusetts birth

certificate data censor any race-name observations with fewer than five occurrences, so there are

fewer than five annual occurrences among African Americans for all of our white names. The

8

first names we use to represent African American home-seekers all have a relative liklihood

measure of at least 44.5 (four are infinite). We also used the same last names as Bertrand and

Mullainathan (2004). Table 1 shows a list of all the names and their frequency of use in the

experiment.

One of the difficulties5 with using on-line audits is that landlords may infer something

other than race from an auditor’s name that may influence their decision to respond. One

concern with using names to identify race is that landlords may infer more than race from a

home-seeker’s name.6 Specifically, they may also infer the social class of the individual. This is

particularly problematic for studying racial discrimination using names if African American

sounding names are associated with a lower class, as this would bias the results toward finding

racial discrimination. We confront this problem by introducing a notion of social class directly

into the experiment.7

We introduce social class by contacting landlords using several high-class and low-class

e-mail messages (see appendix for examples of each type). Both the high-class and low-class e-

mail messages express interest in a rental property advertised on-line, and both offer to send

references and credit information upon request. The high-class and low-class e-mail messages

differ by how information is conveyed. Low-class e-mails all contain a spelling error, informal

5 We address several other possible confounding influences on landlords, including name association with religious beliefs, and name uniqueness in Section VI. 6 If landlords do not infer anything about a home-seeker from their name, we would expect our experiments to show no difference in how inquires from the African American and white sounding names in our sample are treated. 7 The notion of social class is often used, but generally difficult to precisely define or quantify. In our experiments, we intend for landlords to infer information (besides race) specific enough for them to believe an individual would be a better (high-class) or worse (low-class) tenant for the property, yet vague enough not to be associated with a particular characteristic such as income, occupation, or education. We apply the notion of social class described by Milton Gordon as, “…class has no precise, agreed-upon meaning but is used either as an omnibus term, to designate differences based on wealth, income, occupation, status, group identification, level of consumption, and family background, or by some particular researcher or theorist as resting specifically on some one of these enumerated factors. There is substantial agreement, however, that the stratifications of class are not by definition those of race, religion, and ethnic origin,” Gordon (1949), p.1.

9

or grammatically incorrect sentences, and an informal valediction before the e-mail signature.

High-class e-mails suggest that the references they have are “good” and the credit report is

“recent”.

The design of our experiments reflects differences in both the home-seekers’ race

(inferred from the name) and class (inferred from the type of e-mail). The experiments test the

effect of race and class separately, as well as jointly. To do this, we randomly assign a pair of e-

mails to send a landlord from the following matrix:

Matrix 1: Audit Types

White African American

High Class Type 1 Type 2

Low Class Type 3 Type 4

In our design, landlords receive a pair of e-mail inquiries from one of the six combinations of

types shown in Matrix 1.

Each landlord receives exactly two e-mail inquiries. For example, one landlord may

receive an inquiry from a Type 1 home-seeker (white name and high-class e-mail) and a Type 2

home-seeker (African American name and high-class). In this example, the experiment tests

only the difference in race. We isolate the effects of race and class by performing audits where

the inquires only differ by class (landlords receive either a Type 2 and Type 4 inquiry, or a Type

1 and Type 3 inquiry). We estimate the joint effect of race and class by performing audits where

the inquires differ by both race and class (landlords receive either a Type 3 and Type 2 inquiry,

or a Type 1 and Type 4 inquiry).

10

One immediate concern with the audit-style design of e-mailed experiments is the order

of sending inquiries. We control for any ordering effects by sending out all six audit

combinations from Matrix 1 both with the white name sending the first e-mail and with the

African American name sending the first e-mail. This doubles the total number of audit types to

12, each listed in Table 2. Sending each type of audit with each race sending the e-mail first

ensures that we have an approximately equal number of audits where an e-mail is sent first from

the white name and from the African American name. Randomly assigning among the set of

audit types ensures that order of sending e-mails does not confound our results.

The venue for our experiments is the popular classified advertisement website Craigslist

(www.craigslist.org). Craigslist allows participants to place and reply to on-line advertisements

specific to local markets for jobs, rental housing, companionship, and other goods and services.

We use only the listings pertaining to the rental housing market. Craigslist is widely used; it has

more than 50 million unique U.S. visitors, and receives more than 20 billion page views

monthly.8 Landlords may create an advertisement for their unit at no monetary cost, and home-

seekers may reply to an unlimited number of advertisements at no monetary cost. 9

Using an on-line venue for discrimination experiments has several advantages compared

to those performed by human actors. The on-line venue ensures that the same landlord receives

the same inquiry, and that it is the delivered in the same manner from both auditors. An on-line

venue also provides a cost-effective way to increase the sample size of the experiment.

8 According to http://www.craigslist.org/about/factsheet 9 Craigslist puts a notice on every advertisement posted on the site warning that stating a discriminatory preference is illegal and allows users to easily notify them of any such advertisements.

11

To implement the experiment we first need to create a set of landlords that have

properties available for rent. The sample starts with the universe of advertisements10 posted on

Craigslist for a given local market made between 9 A.M. and 9 P.M. on Tuesdays. The

experiment uses advertisements posted every Tuesday from July 22, 2009 to October 7, 2009

from the Atlanta, Boston, Chicago, Dallas, Washington, D.C, Houston, Los Angeles, New York,

Seattle, and San Francisco local Craigslist pages.11 We remove repeat landlords by manually

scanning each posting and selecting only unique housing units. First, we do this visually by

eliminating postings made for the same rental property, and then automatically remove any

postings that contain the same landlord contact information.

We randomly select a sample of rental units from the list of unique posted

advertisements, and then randomly match each unit with an audit-type.12 After randomly

assigning an audit-type (see Table 2), each posting is randomly matched with the appropriate

home-seeker names (signifying race) and e-mail copy (signifying class). We also match to

ensure that in the case of a white-white or African American-African American audit we do not

send inquires from the same name. We conduct the random assignment of audit-type so that we

send the same number of e-mails from each audit-type for each city/day.13

10 Craigslist identifies each advertisement with a unique posting identification number. We use this number to identify advertisements to match unit characteristics and responses. 11 The Boston and New York craigslist pages separate housing advertisements between those made by a broker that charges a fee and those that do not. For our experiment, we use only the advertisements listed under the no-fee pages in these cities. 12 In most cases, the random sample is 252 postings for each city (a large number divisible by 12 audit types). Because of time and resource constraints, in most weeks we chose two cities to draw a sample from. The exception to choosing at least 252 postings is in the Dallas and Houston markets, because the number of unique postings is small. Also, we over-sample New York City in one week because of the large universe of postings (504 postings total postings: 252 from Manhattan, Staten Island, and Northern New Jersey and 252 from Brooklyn, Queens, the Bronx, and Long Island). 13 The exception to this rule is combining the Dallas and Houston markets because of small sample size.

12

We send all e-mail inquiries on the Wednesday after the landlord posts the advertisement.

We send all inquiries from g-mail account addresses in the following format:

firstname.lastname.###@gmail.com, where ### is a three-digit number unique to each name.

We send e-mail inquires manually according to name of sender, type of e-mail, and order

assigned randomly through the audit-type.14 We reveal the name (race) of the home-seeker to

the landlord in three different ways: from the e-mail address, from the name-plate in the

landlords’ inbox, and in the signature of each e-mail. Each Wednesday we begin sending the

first wave of e-mails at 9 A.M and finish with the second wave no later than 12 P.M. Due to

concerns of exposing the experiment to landlords, we leave at least one hour between all first and

second e-mail inquiries, and at most three hours.

IV. Sample Characteristics of Units and Surrounding Neighborhoods

In total, our experiment consists of 4,728 audits, or 9,456 e-mail inquiries. The overall

response rate is 53.9 percent, with 63.7 percent of landlords responding to at least one e-mail

inquiry from a given audit pair. 15 Table 3 details the number of audits carried out in each city in

our sample, as well as the overall response rate and percentage of landlords who reply to at least

one e-mail inquiry. The highest overall response rate is 61.9 percent in the Washington, D.C.

area, the lowest is 40.9 percent in the Houston area.16

14 We send e-mail inquires manually because of problems with spam filters in g-mail and with the anonymous e-mail addresses used by craigslist posters that we encountered in trial experiments. 15 Our response rate is comparable to recent studies of discrimination. Ahmed and Hammarstedt (2008) have a response rate of between 21 and 56 percent, depending on the race and gender of the individual, using internet housing advertisements in Sweden. Using only advertisements from Los Angeles, Carpusor and Loges (2006) have a 73 percent response rate. Ewens, Tomlin, and Wang (2009) have a response rate of 64 percent, but do not use New York, and include several smaller markets that we do not. 16 The response rate in a city has only a modest, and not statistically different than zero, correlation with metropolitan area vacancy rates as reported in the 3rd quarter of 2009 by the Housing Vacancy Survey. The response rate in a city is also only modestly correlated with net discrimination, and again not statistically different than zero.

13

Of the 4,728 rental units in our sample, we are able to obtain information about the

characteristics of 3,906 of them.17 There are no standard information fields on Craigslist, so the

information supplied about the unit for rent is up to the discretion of the person placing the

advertisement. Table 4 shows the average characteristics for rental units in our entire sample and

for each city individually. On average, units in our sample have slightly more than 2 bedrooms

and 1.6 bathrooms, and landlords most often describe them as an apartment (47 percent). We

also have a substantial share of units described as single family homes (27 percent), as well as

some described as shared rooms (3.8 percent), duplexes (3.8 percent), and townhomes (6.9

percent).18 The average monthly rent is $1,492, and just fewer than 46 percent of the units in our

sample rent for more than the median19 in the city where they are located. The units in our

sample average 1,361 square feet, and about 15 percent of them offer some type of discount.20

Many of the advertisements in our sample include the address of the rental unit, or a link

providing a map of the unit’s location. We use the location information provided in the

advertisement to match each rental unit to a census tract.21 A census tract match is valuable

because the census maintains detailed information about the residents of tracts so that we can 17 We are not able to obtain the characteristics of all rental units in our sample because recording this information is time sensitive. Craigslist deletes some advertisements because of complaints by users; landlords may also delete advertisements (possibly, for instance, if the property rents quickly). We did not want to bias our experiments by purposefully excluding these types of advertisers so we created our sample using all units posted on a given day. 18 The unit-type categories do not add to 100 percent because categories are not necessarily mutually exclusive. We categorized units according to the description given by the advertisement. For example, the unit advertised as a shared room in a single family home would appear in both of these categories in our data. In addition, not all advertisements describe the type of unit, so there are units that we have some data on (such as the rent) but are not able to determine the type. 19 Median rents are from fiscal year 2010 estimates by the Department of Housing and Urban Development (HUD). We match a unique median rent to each unit by city and number of bedrooms of the unit to determine if the unit rents for above or below the city median. 20 The typical discounts offered are reduced rent for an introductory period, a lower security deposit, and including utilities with the advertised rent. The discount variable does not differentiate the type or value associated with a discount. 21 The Census defines tracts as small, relatively permanent statistical subdivisions of a county that usually have between 2,500 and 8,000 persons designed to be homogeneous with respect to population characteristics, economic status, and living conditions. Tracts are entirely contained by counties and the size (in terms of land area) varies widely depending on population density (www.census.gov).

14

determine the characteristics of the surrounding neighborhood for rental units.22 Of the 3,906

units that we have some characteristic information on, we are able to match 3,060 to a census

tract and obtain information on the surrounding neighborhood.

Table 4 shows the average neighborhood characteristics for rental units for our entire

sample, and for each city individually. On average, units in our sample are in neighborhoods

with 68 percent white residents and 14 percent African American residents. The standard

deviation on both the percent of white residents (.2535) and the percent of African American

residents (.2286) is substantial, indicating that we have neighborhoods in our sample that range

from vast majority-white to majority-African American. On average, 13 percent of residents in

the neighborhoods surrounding our units live below the poverty level, while median family

income is $65,134. In addition, 36.5 percent of residents in neighborhoods surrounding the units

in our sample have at least a college education.

Across cities, there are stark differences in the racial composition of neighborhoods

surrounding the units in our sample. In Atlanta (31.7 percent) and Washington, D.C. (27.5

percent), the average surrounding neighborhood has a high percentage of African American

residents relative to other cities. In Boston (5.7 percent), Seattle (4.8 percent), and San Francisco

(6.8 percent), the average surrounding neighborhood has less than 10 percent African American

residents. Surrounding neighborhoods in New York (57 percent) and Los Angeles (59.2 percent)

have the lowest percentage of white residents and a low percent of African American residents

(20 percent in New York and 10 percent in Los Angeles), indicating that those neighborhoods

have a substantial share of residents that are a race other than African American or white. 22 We match the units in our sample to census tracts by the address of the unit or the address of the nearest intersection (if a link to a map is provided) using the census American FactFinder Address Search. This search allows the user to input a street address, city, state, and zip code and outputs the associated census tract. The address search is available on-line at http://factfinder.census.gov.

15

V. Results

Table 5 shows the results of landlord response to our inquiries about advertised rental

housing. Columns (1)-(3) of Table 5 present the results by looking at the response rate by race

for all e-mails (9,456). Columns (4)-(8) of Table 5 show the results at the audit level using only

audits where landlords receive e-mails from both an African American and a white home-seeker

(3,153 audits). The response rate for African American home-seekers is 6.3 percentage points

lower than for whites, statistically significant at less than the one percent level.23 Across most

cities in our sample, the response rate for African Americans is between 4 and 6 percentage

points lower than whites. In Los Angeles and Boston, the response rate difference is

significantly larger than in all other cities- especially interesting as these cities were the only

ones used in prior studies by Yinger (1986) and Carpusor and Loges (2006), respectively.

Nearly 80 percent of landlords in the experiment either respond to both inquiries or to

neither inquiry, as shown in Columns (4) and (5) of Table 5. Although we count these landlords

to represent equal treatment, it is possible that they do not infer anything about race from the

names we use. Columns (6) and (7) of Table 5 present the proportion of landlords that respond

differently to white or African American home-seekers, respectively. For the full sample, 12.5

percent of landlords respond only to the inquiry from a white sounding name, while about 8

percent respond only to an African American sounding name. Column (8) shows the p-value for

a difference in proportions test between Column (6) and (7), which measures the net incidence of

discrimination at the landlord level. For the full sample, we reject the null hypothesis that the

23 All p-values reported in the tables are generated from a z-test and reflect the null hypothesis that the difference in proportions between two independent groups equals zero. A p-value of 0.0001 means that we easily reject that the difference in proportions equals zero.

16

proportion responding only to white home-seekers equals the proportion responding to only

African American home-seekers.

The level of discrimination we find is generally lower than what other studies using an

internet venue report. Carpusor and Loges (2006) find a response rate difference of 33

percentage points using rental properties posted on craigslist for the Los Angeles market.

Ewens, Tomlin, and Wang (2009) also report higher levels of discrimination using rental

properties posted on craigslist– almost a 10 percentage point lower response rate for African

American males (although the difference shrinks to 5.9 percentage points for females). In

addition, Ahmed and Hammarstedt (2008) report net discrimination of 24.8% favoring Swedish

male names over Arabic/Muslim male names.

Although our outcome (response) is not directly comparable to most studies that use in-

person audits, we find lower levels of discrimination than most of these studies. Ondrich,

Stricker, and Yinger (1998) find the probability of receiving a call back from real estate agents is

4.6 percentage points lower for African Americans. Yinger (1998) reports net discrimination of

7.7% for follow-up calls in sales audits and a net discrimination of 10.7% for exclusion from

available units in the rental audits. Additionally, Zhao (2005) finds that real estate agents show

African Americans 30% fewer homes than whites. It is difficult to say if the smaller magnitude

of our results is a function of removing the bias from in-person audits, or other factors such as

using the internet for search or the fact that our research takes place nearly 30 years24 after some

of these studies.

Response by Race and Class

24 Yinger (1986) studies audits conducted in Boston in 1981.

17

Table 6 summarizes the response to inquiries for advertised rental housing units by race

and class of home-seekers. The low-class e-mails from white sounding names have the highest

response rate at 57.6 percent, while the low-class e-mails from African American sounding

names have the lowest response rate at slightly more than 49 percent. Comparing the response

rates shows that high-class e-mails from African American sounding names perform better than

low-class e-mails from the same set of names, but still not as well as white sounding names from

either class type. The pair that receives equal treatment least often is when a landlord receives a

high-class e-mail from a white sounding name and a low-class e-mail from an African American

sounding name (77.3 percent of the time). We find strong evidence of discrimination against

African American home-seekers when both inquiries are made using low-class e-mails. 25

When a landlord receives e-mail inquiries that are both high-class, they are only slightly

more likely to respond to only the white sounding name (11.27 percent vs. 9.49 percent).

Furthermore, we cannot reject the null hypothesis that the net level of discrimination between

African American and white home-seekers equals zero when both inquiries use a high-class e-

mail. Although overall, we do not find statistically significant net discrimination when we send

high-class e-mails from African American and white home-seekers, Panel (1) of Figure 1 shows

that in Los Angeles there is statistically significant net discrimination for this pairing. When

landlords receive a low-class e-mail from both races, they tend to only respond to the home-

seeker with a white sounding name; however, as shown in Panel (6) of Figure 1, this difference

is only statistically significant in Boston, Los Angeles, and Washington, D.C.

25 Due to the lower number of observations, the power of these tests is lower than tests using the full sample. We are able to reject a difference of about 3 percentage points between groups, well within the full sample estimate of 4.5 percentage point difference between groups.

18

Another way of demonstrating the marginal effects of race and class and their

interactions on the likelihood of landlord response is to examine the coefficients of probit

regressions at the e-mail level. In these regressions the dependent variable equals one if the e-

mail receives a response, and zero if it does not. We represent race of the sender with a dummy

variable equal to one if we send the e-mail from an African American sounding name. We

represent class of the sender with a dummy variable equal to one if we send the e-mail using a

low-class e-mail, and we use a dummy variable for order that equals one when we send the e-

mail second.

Table 7 shows the marginal effects calculated from probit regressions. We calculate all

p-values shown in Table 7 from standard errors clustered at the landlord level. The marginal

effects in Columns (3), (6), and (9) of Table 7 include landlord fixed effects and correspond to

the landlord response differences in Tables 5 and 6. The results with landlord fixed effects show

a much larger marginal effect of race than the response rate results. The size of the coefficient is

larger and the number of observations is smaller in these regressions because they identify the

marginal effect of race on response from landlords who only reply to one e-mail (landlords that

reply to both or neither e-mail perfectly predict the outcome and are dropped from the probit).

The marginal effect of an African American name (-0.222) on response is exactly equivalent to

the difference in response rates for whites (394/645, or 61 percent) and African Americans

(251/645 or 39 percent) for landlords that only reply to one e-mail shown in Table 5. As with the

landlord level results these results are not subject to omitted variables at the unit, landlord, or

neighborhood level that may plague the results that do not use fixed effects.

Columns (1), (4), and (7) of Table 7 show marginal effects from probit regressions that

do not control for any observable unit or neighborhood characteristics in our dataset. Columns

19

(2), (5), and (8) show marginal effects from probit regressions using the full set of control

variables, 26 thus the number of observations is substantially smaller. The marginal effects

without landlord fixed effects correspond to the response rate differences shown in Tables 5 and

6. Column (4) of Table 7 shows that the marginal effect of race, after removing the effect of

class and the joint effect of race and class, is somewhat smaller than the effect shown in Column

(1) without controlling for class. In Column (4), the marginal effect of sending an e-mail from

an African American sounding name is a 4 percentage point reduction in the likelihood of

response, statistically significant at conventional levels.

The marginal effect of race shrinks further when we control for the order of the e-mail

inquiry and the interactions between race, order, and class. Column (7) shows that when we

control for class, order, and the interaction between them and race, the marginal effect of sending

an e-mail from an African American sounding name is a 3.6 percentage point reduction in the

likelihood of response, with only marginal statistical significance (p-value of 0.077). The results

in Columns (2), (5), and (8) that control for the full set of information on the unit and

neighborhood also reduce or eliminate the importance of race; however, these results rely on a

small sub-sample of the data- only 830 observations.

Response by Neighborhood and Unit Characteristics

Using the information on unit characteristics from the posted advertisements and the

neighborhood from our census tract match, we can test for differences in discrimination across

26 We use the following control variables in these models: an indicator of the city where the local craigslist advertisement is placed (we exclude the indicator for Seattle), number of bedrooms and bathrooms, the square footage of the unit, if the unit is an apartment or single family home, monthly rent, an indicator if the monthly rent is above the median, the percentage of residents living below the poverty line, the percentage of residents with at least a college degree, the percentage of white residents, and an indicator if the neighborhood is a "tipping" neighborhood (between 80 and 95 percent white).

20

characteristics and examine how they may impact discrimination on the margin. Looking across

characteristics can more accurately demonstrate where discrimination occurs. It also offers some

insight as to why discrimination occurs.

Card, Mas, and Rothstein (2008), suggest that the absolute share of minorities in an area

has a substantial effect on out-migration of white residents. They suggest that neighborhoods

exhibit tipping behavior when the minority share is between five and 20 percent- at that point

white residents move almost entirely out of neighborhoods. If this is true, and landlords want to

prevent neighborhood tipping, they may be more likely to discriminate in neighborhoods where

the share of non-whites is between five and 20 percent, than in other neighborhoods. Other

researchers have tested for differential discrimination in tipping neighborhoods including Page

(1995) who finds a sharp increase in discrimination for neighborhoods that are exactly 20 percent

African American, and Yinger (1986) who finds differential discrimination by racial composition

for neighborhoods in Boston.

We augment the test of discrimination to determine if landlords of properties in

neighborhoods with between five and 20 percent minority residents discriminate more than in

other areas. Table 8 shows the results, augmented by the absolute percentage of white residents,

support the idea that landlords discriminate more in neighborhoods that are in the five to 20

percent tipping point range described by Card, Mas, and Rothstein (2008). For both the response

rate and landlord level tests, we find that discrimination is most severe in neighborhoods that are

between 80 and 95 percent white.27

27 Similar to Page (1995), we also test for sharp changes in discrimination across neighborhood racial composition by estimating the following probit model,

where Y is a zero-one dummy variable indicating response or non-response to our e-mail inquiry, “pct non”

21

Discrimination is also sensitive to the type of unit advertised on Craigslist, as shown in

Table 9. We find no evidence of either response differences or landlord net discrimination for

units advertised as single family homes, duplexes, or townhomes. We find strong evidence, both

in response rate difference and landlord net discrimination, for units advertised as condominiums

or apartments. The difference in discrimination between the different types of units may reflect

landlords’ beliefs about current or future tenant prejudice.28 If landlords advertising apartments

or condominiums are concerned with being able to rent additional units in the same building and

they believe current or future tenants are prejudice, then they may be more likely to discriminate.

Landlords advertising free standing units (or units with fewer tenants) may be less concerned

with renting additional units, and therefore may not take preferences of current or future

neighbors into account.

To more fully understand the effect of unit and neighborhood characteristics on

discrimination we estimate probit regressions at the e-mail level, where the dependent variable

equals one if the landlord replied, and zero if they did not. Each regression includes a dummy

variable indicating if the sender has an African American sounding name, the characteristic of

interest, and an interaction between the African American name indicator and the characteristic

of interest. Table 10 shows the marginal effects29 of the interaction between the African

American sounding name indicator and each characteristic listed in the first column.

measures the percentage of non-white residents in a neighborhood, and “AA” indicates that the e-mail is sent from an African American name. We estimate this model for X equal to 5, 10, 15, 20, 25, and 30 percent, respectively. Only the results (available upon request from the authors) with X equal to 10 percent non-white show a statistically significant jump in discrimination. 28 This is the customer-prejudice hypothesis; see Ondrich, Stricker, and Yinger (1998) and Zhao (2008) for a more complete explanation and additional evidence. 29 We calculate marginal effects at the mean value of the dependent variable. Marginal effects reflect the percentage point change in the probability of response given a 1 unit change in the characteristic, or going from the absence of the characteristic to the presence of the characteristic for indicator variables.

22

Table 10 reveals that the only statistically significant predictors of landlord response

(besides race, class, and order of e-mail inquiry) are the size of the unit, the number of bedrooms,

whether the unit is an apartment or single family home, and if the unit is in a neighborhood

identified as a “tipping” neighborhood as defined by Card, Mas, and Rothstein (2008). The

number of bedrooms, the square footage of the unit, and listing the unit a single family home

have a positive marginal effect on response for e-mails we send from African American

sounding names. Being in an apartment building and being located in a “tipping” neighborhood

has a negative marginal effect on the likelihood of response for e-mails we send from African

American sounding names.

Intensity of Response

We find evidence that landlords discriminate by responding more often or only to e-mail

inquiries we send from home-seekers with white sounding names. It is true, however, that nearly

40 percent of landlords in our experiment respond to both the white and African American

sounding names. Responding to both inquiries does not necessarily mean that landlords treat

each one the same. To determine if landlords that respond to inquiries from white and African

American sounding names treat those inquires the same we examine the intensity of each

response. We measure the intensity of a response by whether the landlord describes the unit as

available, and whether the landlord invites the home-seeker to view the unit.

Table 11 shows the results of testing for net discrimination in terms of whether the

landlord described the unit as available.30 Nearly 69 percent of landlords who respond mention

that the unit is still available to both groups, while almost 20 percent did not mention availability

30 We use only the experiments where a landlord replied to e-mail inquiries from both an African American and white home-seeker; this is 1,356 landlords.

23

to either group. The net difference in landlords describing the unit as available between whites

and African Americans is not statistically different that zero for the entire sample, or in any of

the cities in our sample individually.

Table 12 shows the results measuring net discrimination by the difference in whether the

landlord invited the home-seeker to view the unit. To avoid double-counting discriminating

landlords, we use only the sample of landlords that describe the unit as available to both inquiries

for these tests (935 audits). More than 58 percent of landlords who described the unit as

available invited both home-seekers for a showing, while nearly 28 percent invited neither for a

showing. As shown in Column 5, we cannot reject the null hypothesis that there is no net

discrimination in inviting home-seekers to view the unit for the full sample and almost all cities

individually- the exception being Seattle where African Americans are invited to view more

often than whites.

VI. Robustness of Primary Findings

The audit style of our experiment removes many of the factors that confound identifying

racial discrimination. Most importantly, it controls for any fixed factors about the landlord. We

also eliminate bias that may arise with in-person audits by conducting our experiment via e-mail

correspondence. There are, however, still a few important factors in our experiment that suggest

some robustness checks.

One choice we have not justified is the use of last names associated with each race. The

last names we choose are from Bertrand and Mullainathan (2004), who offer no justification for

last name choice other than that the last names sound like they are more likely to be associated

with whites or African Americans. To test how much the choice of last name matters, we

24

purposefully used the last name Jones associated with both an African American (Rasheed) and a

white (Todd) first name in our experiments. The overall response rate for Todd Jones is 58

percent, while the overall response rate for Rasheed Jones is 49.4 percent, a difference of 8.6

percentage points, statistically significant at the one percent level.31 The difference based only

on first name choice is only slightly larger than the full sample difference; we therefore conclude

that last names are not a decisive factor in our experiment.

There are three other concerns with using the first names from Bertrand and Mullainathan

(2004) that we address here. The first is that some of the African American names are not only

associated with race, but are also distinctly of Muslim origin- Hakim, Jamal, Kareem, and

Rasheed are all linked to Muslim origin.32 The second is that our names may not only be

associated with race, but also with uniqueness- and landlords may treat unfamiliar names

differently. The third concern is the names come from birth certificate data in Massachusetts,

and our experiment is nation-wide. If the names we use are more/less common across the cities

in our sample, or more/less likely to be associated with only one race, our cross-city comparisons

may not be valid.

To address the concern that a Muslim association with some of our African American

names may be driving our results we exclude e-mails sent from the names Hakim, Jamal,

Kareem, and Rasheed and re-estimate our primary results. The first row of Table 13 shows our

re-estimates excluding these names using the response rate (excluding all e-mails from these

names) and landlord level (excluding all audits that use these names). These results suggest a

slightly higher level of discrimination towards African Americans than the full sample results. 31 The number of Todd Jones paired with Rasheed Jones audits is only 36. Of those 36 audits, 16 landlords replied to both e-mail inquiries, 15 landlords replied to neither, 3 replied only to Rasheed Jones, and 2 replied only to Todd Jones. The difference in landlord level response is not statistically different that zero. 32 We thank a particularly thoughtful referee for pointing this fact out.

25

At the landlord level, net discrimination happens from almost 6 percent of landlords, as opposed

to 4.5 percent in the full sample. At the e-mail level, African Americans have a lower response

rate by 7.2 percentage points, as opposed to only 6.3 percentage points using the full sample.

To address concerns with the uniqueness of names in our sample, we use data from the

Social Security Administration on names of babies born during the years 1974-1975 (following

the Bertrand and Mullainathan birth certificate data). The data show the most common names

nationally and for all the states in our sample. Table 14 presents the average ranking for all of

the names we use in our experiment. As shown in the first column, all of the names in our

sample appear in the 1,000 most common male names (these ranking are for the full population,

not particular to any race). Not surprisingly, almost all of the white names in our sample are

more common than African American names- the exceptions being Tyrone and Jermaine. Table

14 also shows the popularity ranking by state in our sample (city level data is not available) for

the names in the top 100. Again, in most cases the white names are far more common that the

African American names, which are often not in the 100 most popular names.

We re-estimate our primary results excluding several combinations of common and

unique names, shown in Table 13. First, we exclude both the two most common (Matthew and

Todd) and two most unique (Hakim and Rasheed) names and re-estimate. Row 2 of Table 13

shows that this restriction does not change our primary results. We further restrict the sample to

exclude the four most common (Matthew, Todd, Brett, Brad) and four most unique (Rasheed

Hakim, Tremayne, Kareem) names and find a reduction in our measure of discrimination.

Excluding these names reduces measured discrimination at the landlord level by 1.5 percentage

points and by almost a half percentage point in the response rate difference.

26

We further test how these names drive our results by examining only the sample of audits

that uses either one of the four most common the four most unique names, shown in Row 4 of

Table 13. These results show a higher level of measured discrimination than the full sample

results- almost 5.5 percent of landlords as opposed to 4.5 percent in the full sample, and a

response rate difference of more than 7.5 percent as opposed to 6.3 percent in the full sample.

As shown in Table 14, the uniqueness of the names we use differs across the geographic

areas of our study. As a test of how sensitive our results are to name uniqueness by location, we

re-estimate our results excluding either names among the 50 most common (Row 5 of Table 13)

or 100 most common (Row 6 of Table 13) by state level location. Excluding names among the

50 most common shows a slight increase in discrimination at the e-mail level, and a slight

decline at the landlord level. Excluding names among the 100 most common causes a drop of

more than 1 percentage point in the response rate difference, but almost no change in

discrimination measured at the landlord level.

On balance, these results suggest that uniqueness of names may play a role in the

magnitude of our findings, but that discrimination still exists even when excluding the most

unique and most common names. Furthermore, the African American names we use, while less

common for the general population, are still among the most common names for African

Americans, and are therefore representative of the actual treatment an African American using

Craigslist could expect.

A third concern with the names we use is that the names come from birth certificate data

in Massachusetts, and our experiment is nation-wide. If the names we use are less (more)

common across the cities in our sample, or less (more) likely to be associated with only one race,

27

our cross-city comparisons may not be valid. We compare the response rate by name and city

relative to the response rate by race and city, to see if certain names garner a differential response

in some cities.33 Table 15 reports p-values for tests of symmetry between the response rate of a

particular name/city combination and the response rate of all names in that race/city

combination. For example, Row (1), Column (1) of Table 15 is the p-value for a test of

symmetry between the response rates for e-mails sent from Brad Davis to advertisements made

in Atlanta and e-mails sent from all other white sounding names to advertisements made in

Atlanta. In most cases, we cannot reject the null hypothesis that the response rate between any

one name/city combination and the corresponding race/city combination are equivalent. As

Table 15 shows, there are cases where a particular name from either race did better or worse than

other names of the same race for that city.

We test if the performance of particular names is driving our results by excluding all

statistically different name/city combinations.34 Table 16 shows that excluding unusual

performing names from the sample does not change the overall results, we find only a slightly

lower response rate gap and landlord level difference. Measuring discrimination by the response

rate difference, we find more discrimination in Atlanta (now statistically significant), less

discrimination in Boston and Los Angeles (still statistically significant), and less discrimination

in Houston and San Francisco (still not statistically significant). Using the landlord level

33 Note that we are not comparing the same name across cities, which would pick up the difference in discrimination that exists between those two cities. Our comparison picks up any differential treatment within the city and race across names. 34 We exclude all e-mails for the response rate difference and all audits where either e-mail contains one of the unusual performing names. For this robustness check, we exclude results for Leroy Parker in Atlanta, Tyrone Cooper in Boston, Brett Murphy in Dallas, Rasheed Jones in Dallas, Tyrone Cooper in Houston, Jamal Robinson in Los Angeles, and Jermaine Jackson in San Francisco.

28

response, we cannot reject no net discrimination in Houston, and lose the marginal significance

of net discrimination in San Francisco.35

External Validity Concerns

Although the audit style experiment we conduct using e-mail inquiries eliminates most

concerns about the internal validity of our results, there are some concerns about how valid our

results are outside of the set of landlords and rental units posted on craigslist. We can think of a

few concerns about the external validity of our results.

The first concern about external validity is that we only respond to one posting per

landlord, even though many landlords make multiple advertisements (both for the same unit and

for other units they own or manage). As part of our experimental design, we deliberately

attempted not to inquire about rental housing from the same landlord in more than one audit. We

decided to only include one audit for repeat landlords in our experiment after running pilot

experiments in the Nashville, TN and Philadelphia, PA markets. During these pilots, we found

some landlords mentioned receiving multiple inquires from the same set of names and seemed

suspicious of this. We saw this as a tradeoff between accurately identifying discrimination by

not exposing the experiment and measuring a more representative level of discrimination.

Unfortunately, we do not know how many postings (or how many units) these landlords make,

so we cannot assign weights to our data. If landlords with multiple units listed on craigslist are

35 We also test our Table 5 results excluding the names in cities that have a different response rate than other names from the same city/race group at the ten percent level. For these results, the overall level of discrimination increases by .5 percentage points in response rate difference and the net level of discrimination at the landlord level increases by .13 percentage points. Atlanta, Washington, D.C, Los Angeles, and New York all show slightly higher levels of discrimination, while San Francisco shows less discrimination and the significance at the 10 percent level goes away.

29

more(less) likely to discriminate than others this means we under(over) estimate the true level of

discrimination.

Using the internet to identify racial discrimination may also affect the external validity of

our results because landlords can interact with home-seekers anonymously. 36 Our data provides

an opportunity to test if discrimination is more severe when landlords maintain a higher degree

of anonymity because Craigslist gives landlords the option to use an anonymous e-mail address.

The craigslist e-mail address offers no identifying information to home-seekers, and can be set

up to forward inquires to an e-mail address of the landlords choosing.

Table 17 shows the results of our experiment for only landlords that use the anonymous

craigslist e-mail address instead of a personal or professional e-mail. Using only anonymous e-

mail addresses we find slightly less (0.39 percentage points) net discrimination at the landlord

level, and slightly more (0.08 percentage points) discrimination in response rates.37 The biggest

difference using anonymous e-mail addresses is that discrimination is no longer evident in the

Washington, D.C. market.

Another concern with how our results translate to the housing market outside of rental

units advertised on craigslist is how similar the landlords, housing units, and neighborhoods are

to the larger market. Ideally, we would be able to compare the characteristics of landlords that

use craigslist to landlords that do not; however, this is not possible as craigslist does not collect

this information- anyone can post an advertisement without approval, and anyone can reply to a

36 An alternative explanation as to why anonymity matters is that landlords who are more likely to discriminate select into advertising on craigslist (as opposed to other venues) because of the anonymous nature of contact with home seekers. Either explanation would result in correlation between discrimination and landlord anonymity. 37 We also separately test for discrimination among the sample of landlords that use a non-anonymous e-mail address. These results show slightly more discrimination at the landlord level (1.5 percentage points) and slightly less discrimination in response rates (.35 percentage points) than the full sample.

30

posting without registering. To get an idea of how representative the housing units and

neighborhoods in our sample are, we compare characteristics of our rental units to characteristics

of rental units in the metropolitan areas we use sampled by the American Housing Survey,38 and

characteristics of our neighborhoods with all neighborhoods in our cities using census tract data

from the 2000 Census.

Table 18 shows the unit and neighborhood characteristics for our sample compared with

AHS and Census data. Table 18 shows that, in general our craigslist sample varies considerably

from the AHS sample of housing units. Our units are larger and have more bedrooms and

bathrooms- probably because our sample is much more likely to include single family homes and

much less likely to include apartments. Given the larger size of units in our sample, it is not

surprising that they rent for over $450 more per month than units in the AHS. Table 18 also

shows that the neighborhoods surrounding our units have more white residents, higher median

family income, and a greater percentage of college graduates.

Given our results show landlords of apartments are more likely to discriminate, African

Americans may be subject to more discrimination in searching for housing in the pool of rental

units available outside of craigslist than we find here. If we re-weight our results by unit type to

match the distribution of home types in the AHS survey, putting a larger weight on our

apartment results and a smaller weight on our single family home results, we do in fact find more

discrimination against African Americans. Weighted this way, our results suggest that the

38 The unit of observation in the AHS is the dwelling. The AHS data consists of householder responses to survey questions on the actual dwelling and the composition of the occupants of the dwelling. The homes surveyed in the AHS include a core sample of homes that has not changed since 1985 and newly constructed dwellings added to the core annually by sampling addresses from building permits data. The data contain a wealth of information about the dwelling, including if it is owner-occupied. We use only renter-occupied units to create our sample statistics.

31

response rate difference is 8.48 percent (as opposed to 6.3), and net discrimination occurs in 5.58

percent of landlords (as opposed to 4.54).

Finally, minorities may not use Craigslist to find rental housing so that they would not

find themselves subject to the discrimination we expose in our experiment. We have no data to

test whether minorities are more or less likely to use craigslist than other venues; however, we

feel that this is a reason why in-person audit studies and audit studies using other venues are

necessary to uncover discrimination in the housing market.

VII. Conclusion

Our results show significant discrimination by landlords against e-mail inquires made

from African American names in the rental housing market. Discrimination goes away when

both e-mail inquiries use a high-class e-mail, and is stronger when both e-mail inquiries use a

low-class e-mail. We also find that discrimination is more severe for housing units advertised as

part of a building where the landlord potentially owns multiple units (apartments and

condominiums) and in neighborhoods prone to shifting away from a white-majority.

The difference in discrimination by class of e-mail and across unit and neighborhood

characteristics suggests the source of discrimination may be statistical in nature rather than based

on landlord preferences for discrimination.39 Under statistical discrimination, a landlord

discriminates because they have imperfect information about potential tenants and use

observable characteristics to make inference about the type of tenant the applicant will be.

Statistical discrimination should occur because landlords are maximizing profits under

uncertainty, not because of personal preference. Our results suggest landlords may infer

39 Either type of discrimination is illegal in the United States under the Civil Rights Act of 1968.

32

information about the characteristics of a home-seeker from a racial-sounding name that they

believe useful in determining their merit as a tenant. As we demonstrate, when a landlord is

offered something else to infer characteristics from (how an e-mail is written), and the

information is positive, they do not base their response on the race associated with our inquiries.

We find that landlords advertising units as apartments and condominiums discriminate,

while those advertising single family homes, duplexes, and townhomes do not. If landlords

believe current or future tenants are prejudiced, then statistical discrimination could explain

discrimination in larger buildings. In the case of landlords who operate in apartments or

condominiums, they bear the full costs of having an African American tenant move in if the

other tenants are displeased (neighbors may move out or rents may decrease); while landlords of

single family homes may share the cost with other landlords in the neighborhood. An alternative

explanation could be that landlords who rent out apartment buildings are more likely to practice

taste based discrimination.

It is difficult to truly separate landlord prejudice from perceived customer preference.

Increased discrimination in neighborhoods with a racial composition near a tipping point

suggests landlords appear to have current and future customer preferences in mind when they

discriminate. While discrimination occurs in neighborhoods across the racial composition

distribution, it is strongest in neighborhoods that are close to tipping (between five and 20

percent minority share). This suggests landlords discriminate to prevent white residents from

leaving the neighborhood.

Although we do not present a formal link between discrimination and housing outcomes

between races, it seems reasonable to assume African Americans are less likely to find a housing

33

unit that matches their preferences if they face discrimination. Indeed, Bradburd et al. (2005)

show that in housing markets where search and bargaining take place, racial discrimination can

reduce renter’s consumer surplus by as much as 25 percent (depending on reservation price of

renters and the level of discrimination). Applying the level of discrimination we uncover

(approximately 5 percent of landlords) and assuming a reservation price between 50 and 100

percent of actual rents, the Bradburd et al. results imply that consumer surplus in the rental

market is between 4.5 and 9 percent lower for African Americans than it would be without

discrimination.

Finding discrimination against African Americans in the on-line rental housing market

confirms what the previous literature demonstrates. Our study comes to this conclusion using e-

mail communication, which is not likely subject to the bias that may come from using actors in

an in-person audit. This is especially troubling considering the low bar of compliance and low

cost involved in e-mail communication.

References

Ahmed, Ali, Lina Andersson, and Mats Hammarstedt. 2010. “Can Discrimination in the Housing Market be Reduced by Increasing the Information about the Applicants?” Land Economics, 86(1): 79–90. Ahmed, Ali M. and Mats Hammarstedt. 2008. "Discrimination in the Rental Housing Market: A Field Experiment on the Internet." Journal of Urban Economics, 64(2): 362–72. Altonji, Joseph G. and Rebecca M. Blank. 1999. "Race and Gender in the Labor Market." Handbook of labor economics: 3143–259. Antecol, Heather and Deborah A. Cobb-Clark. 2008. "Racial and Ethnic Discrimination in Local Consumer Markets: Exploiting the Army's Procedures for Matching Personnel to Duty Locations." Journal of Urban Economics, 64(2): 496–509. Bertrand, Marianne and Sendhil Mullainathan. 2004. "Are Emily and Greg More Employable Than Lakisha and Jamal? A Field Experiment on Labor Market Discrimination," American Economic Review, 94(4): 991–1013. Bosch, Mariano, Angeles Carnero, and Lidia Farre. 2010. “Information and Discrimination in the Rental Housing Market: Evidence from a Field Experiment.” Regional Science and Urban Economics, 40(1): 11–19. Bradburd, Ralph, Stephen Sheppard, Joseph Bergeron, Eric Engler, and Evan Gee. 2005. “The Distributional Impact of Housing Discrimination in a Non-Walrasian Setting.” Journal of Housing Economics, 14(2): 61–91. Card, David; Alexandre Mas and Jesse Rothstein. 2008. "Tipping and the Dynamics of Segregation," Quarterly Journal of Economics, 123(1): 177–218. Carpusor, Adrian G. and William E. Loges. 2006. "Rental Discrimination and Ethnicity in Names." Journal of Applied Social Psychology, 36(4): 934–52. Cutler, David M. and Edward L. Glaeser. 1997. "Are Ghettos Good or Bad?" Quarterly Journal of Economics, 112(3): 827–72. Cutler, David M.; Edward L. Glaeser and Jacob L. Vigdor. 1999. "The Rise and Decline of the American Ghetto." The Journal of Political Economy, 107(3): 455–506. Ewens, Michael; Bryan Tomlin and Choon Wang. 2009. "Statistical Discrimination in the U.S. Apartment Rental Market: A Large Sample Field Study." http://dss.ucsd.edu/~btomlin/pdfs/ewenstomlinwang.pdf Heckman, James. 1998. "Detecting Discrimination." Journal of Economic Perspectives, 12(2): 101–16.

Heckman, James and Peter Siegelman. 1993. "The Urban Institute Audit Studies: Their Methods and Findings." In Clear and convincing evidence: Measurement of discrimination in America, ed. Michael E. Fix and Raymond J. Struyk, 187–258. Washington: Urban Institute Press. Milton, M. Gordon. 1949. "Social Class in American Sociology." The American Journal of Sociology, 55(3): 262–68. Myers, Caitlin Knowles. 2004. "Discrimination and Neighborhood Effects: Understanding Racial Differentials in Us Housing Prices." Journal of Urban Economics, 56(2): 279–302. Ondrich, Jan; Stephen L. Ross and John Yinger. 2000. "How Common Is Housing Discrimination? Improving on Traditional Measures." Journal of Urban Economics, 47(3): 470–500. Ondrich, Jan; Stephen Ross and John Yinger. 2003. "Now You See It, Now You Don't: Why Do Real Estate Agents Withhold Available Houses from Black Customers?" The Review of Economics and Statistics, 85(4): 854–73. Ondrich, Jan; Alex Stricker and John Yinger. 1998. "Do Real Estate Brokers Choose to Discriminate? Evidence from the 1989 Housing Discrimination Study." Southern Economic Journal, 64(4): 880–901. Page, Marianne. 1995. "Racial and Ethnic Discrimination in Urban Housing Markets: Evidence from a Recent Audit Study." Journal of Urban Economics, 38(2): 183–206. Ross, Stephen L. and Margery Austin Turner. 2005. "Housing Discrimination in Metropolitan America: Explaining Changes between 1989 and 2000." Social Problems, 52(2): 152–80. Yinger, John. 1986. "Measuring Racial Discrimination with Fair Housing Audits: Caught in the Act," American Economic Review, 76(5): 881–893. Yinger, John. 1995. Closed Doors, Opportunities Lost: The Continuing Costs of Housing Discrimination. New York: Russell Sage Foundation. Yinger, John. 1998. "Evidence on Discrimination in Consumer Markets." Journal of Economic Perspectives, 12(2): 23–40. Zhao, Bo. 2005. "Does the Number of Houses a Broker Shows Depend on a Homeseeker's Race?" Journal of Urban Economics, 57(1): 128–47. Zhao, Bo; Jan Ondrich and John Yinger. 2006. "Why Do Real Estate Brokers Continue to Discriminate? Evidence from the 2000 Housing Discrimination Study." Journal of Urban Economics, 59(3): 394–419.

Appendix: Examples of “high-class” and “low-class” e-mail inquires we send to landlords

High-Class Hi there, I’m interested in the rental you posted on Craigslist, would you tell me if it is still available? If you need them, I have good references and I could also send a recent credit report. Thanks for your time. Sincerely, First and last name of home-seeker

Low-Class

Hi, I saw the place on the internet. Is the place still avialbe? Do you need references or credit scores? I can send those if you want. C U Later, First and last name of home-seeker

White Names Frequency of Occurrence Percentage of e‐mailsBrad Davis 474 5.01Brendan Ryan 513 5.43Brett Murphy 554 5.86Matthew O’Brien 522 5.52Neil Baker 500 5.29Geoffrey McCarthy 526 5.56Todd Jones 529 5.59Greg Young 561 5.93Jay Wright 550 5.82African American NamesDarnell Johnson 557 5.89Hakim Washington 557 5.89Jamal Robinson 503 5.32Jermaine Jackson 549 5.81Kareem Hall 509 5.38Leroy Parker 513 5.43Rasheed Jones 514 5.44Tremayne Williams 539 5.70Tyrone Cooper 486 5.14

Table 1: Names Used in Discrimination Experiment

Notes: The names in our experiment come from Bertrand and Mullainathan (2004) who identify names that have a high likelihood of association with only one race. They determine the likelihood a name is assoicated with a race by examining Massachusetts birth certificate data from 1974‐1979. Two e mails are sent to each landlord in our sample one from each raceTwo e‐mails are sent to each landlord in our sample, one from each race group. All e‐mails were sent to landlords on Wednesday mornings, in response to advertisements posted on Tuesdays. We conducted our experiment from July 22 to October 7, 2009.

Sent First Sent SecondWhite Name/High Class e‐mail African American Name/High Class e‐mailAfrican American Name/High Class e‐mail White Name/High Class e‐mailWhite Name/Low Class e‐mail African American Name/Low Class e‐mailAfrican American Name/Low Class e‐mail White Name/Low Class e‐mailWhite Name/High Class e‐mail African American Name/Low Class e‐mailAfrican American Name/Low Class e‐mail White Name/High Class e‐mailAfrican American Name/High Class e‐mail White Name/Low Class e‐mailWhite Name/Low Class e‐mail African American Name/High Class e‐mailWhite Name/High Class e‐mail White Name/Low Class e‐mailWhite Name/Low Class e‐mail White Name/High Class e‐mailAfrican American Name/High Class e‐mail African American Name/Low Class e‐mailAfrican American Name/Low Class e‐mail African American Name/High Class e‐mail

Table 2: Audit Types

Notes: Audit types reflect all six possible combinations of the elements in Matrix 1, each combination is sent to landlords in standard and reverse order for a total of 12 audit types. We reveal the race of the potential tenant through the name assoicated with the e‐mail inquiry. We reveal the class of the potential tenant through the context of the e‐mail inquiry.

Number of Audits Overall Response Rate Responded to At Least One InquiryFull Sample 4728 53.94% 63.66%Atlanta 504 59.23% 68.85%Boston 504 58.04% 67.86%Chicago 503 49.40% 59.44%Dallas 160 51.56% 63.75%Washington D.C. 504 61.90% 72.42%Houston 296 40.88% 47.97%Los Angeles 492 52.95% 62.60%New York 756 46.56% 57.41%Seattle 504 60.22% 69.05%San Francisco 504 55.46% 64.09%

Table 3: Number of Audits and Response Rate Across Cities

Notes: The number of audits conducted in each city is generally 504, or 252 carried out on two separate occasions. In both Los Angeles (12) and Chicago (1) we had some audits that were unusable because we inadvertently chose duplicate landlords. Dallas and Houston did not have enough unique postings to meetour audit goals individually, so they were combined. We doubled the number of audits in New York in one week (to 504) to take advantage of the large number of postings in that city.

Full Sample Atlanta Boston Chicago Dallas Washington D.C. Houston Los Angeles New York Seattle San FranciscoTotal Housing Units 4728 504 504 503 160 504 296 492 756 504 504Data on Unit Available 3906 456 464 425 150 420 264 445 419 458 405Census Tract Match 3060 349 359 344 98 317 178 382 290 388 355Unit CharacteristicsMultiple Units Listed 0.0220 0.0219 0.0259 0.0141 0.0133 0.0524 0.0417 0.0112 0.0095 0.0262 0.0049

(.1468) (0.1466) (0.1589) (0.1181) (0.1151) (0.2231) (0.2002) (0.1055) (0.0974) (0.1599) (0.0702)Bedrooms 2.04 2.66 1.94 1.88 2.49 2.21 2.10 1.90 1.63 2.22 1.62

( 1.1195) (1.1130) (0.9636) (0.9976) (1.0722) (1.1398) (1.1248) (1.0314) (1.0903) (1.1227) (1.0903)Bathrooms 1.61 1.99 1.32 1.60 1.96 1.88 1.83 1.44 1.21 1.64 1.27

(.7699) (.7414) (0.5975) (0.6491) (0.6074) (0.8407) (0.7665) (0.7746) (0.7106) (0.6712) (0.7023)Single Family Homes 0.2724 0.5744 0.1514 0.0921 0.5248 0.2030 0.3843 0.2921 0.0309 0.4155 0.1898

(.4453) (.4950) (0.3589) (0.2895) (0.5012) (0.4027) (0.4875) (0.4553) (0.1733) (0.4934) (0.3927)Duplex 0.0384 0.0395 0.0314 0.0384 0.0780 0.0198 0.0262 0.0520 0.0254 0.0728 0.0142

( .1922) (.1951) (0.1747) (0.1923) (0.2692) (0.1395) (0.1601) (0.2223) (0.1574) (0.2601) (0.1183)Townhouse 0.0689 0.1119 0.0447 0.0384 0.0567 0.1931 0.0480 0.0644 0.0084 0.0659 0.0227

(.2533) (.3156) (0.2070) (0.1923) (0.2322) (0.3952) (0.2143) (0.2457) (0.0915) (0.2484) (0.1490)Condo 0.1250 0.1349 0.1021 0.1830 0.0638 0.2451 0.1223 0.0891 0.0337 0.1103 0.1136

(.3308) (.3420) (0.3032) (0.3871) (0.2453) (0.4306) (0.3283) (0.2853) (0.1807) (0.3137) (0.3178)Apartment 0.4725 0.1512 0.6990 0.6487 0.2766 0.3102 0.4148 0.4406 0.8596 0.3357 0.5369

(.4993) (.3586) (0.4593) (0.4780) (0.4489) (0.4631) (0.4938) (0.4971) (0.3479) (0.4728) (0.4993)Shared Room 0.0381 0.0138 0.1388 0.0095 0.0000 0.0374 0.1533 0.0137 0.0220 0.0047 0.0077

(.1915 ) (.1169) (0.3461) (0.0972) (0) (0.1900) (0.3609) (0.1165) (0.1467) (0.0681) (0.0874)Monthly Rent $1,492 $1,072 $1,379 $1,390 $1,075 $1,835 $937 $1,633 $1,882 $1,246 $2,071

(811) (416) (499) (583) (587) (837) (454) (943) (917) (535) (1080)Greater than Area Median Rent 0.4594 0.3400 0.3075 0.7214 0.2905 0.5246 0.4458 0.3761 0.5870 0.3208 0.6106

(.4984) (.4742) (0.4620) (0.4488) (0.4556) (0.5000) (0.4981) (0.4850) (0.4930) (0.4673) (0.4882)Square Footage 1361 1649 1154 1298 1417 1820 1398 1442 928 1288 1086

(1209) (956) (397) (545) (701) (3045) (843) (1057) (467) (689) (514)Advertised Discount 0.1537 0.1272 0.2112 0.2476 0.0600 0.1576 0.1742 0.1371 0.1535 0.1444 0.0647

(.3607) (.3336) (0.4086) (0.4322) (0.2383) (0.3648) (0.3800) (0.3443) (0.3609) (0.3519) (0.2463)Neighborhood Characteristics% African American Residents 0.1438 0.3167 0.0565 0.1276 0.1115 0.2752 0.1345 0.1036 0.2098 0.0478 0.0680

(0.2286) (0.3212) (0.1036) (0.2190) (0.1253) (0.3067) (0.1823) (0.1478) (0.2908) (0.0726) (0.1239)% White Residents 0.6847 0.6238 0.8327 0.7252 0.7468 0.6011 0.7184 0.5924 0.5708 0.8225 0.6377

(0.2535) (0.3171) (0.1625) (0.2293) (0.1643) (0.2891) (0.2198) (0.2286) (0.3150) (0.1288) (0.1983)% Residents Under Poverty Line 0.1306 0.1278 0.1156 0.1303 0.1207 0.1098 0.1198 0.1796 0.1832 0.1058 0.1074

(0.1045) (0.1125) (0.0941) (0.1050) (0.0955) (0.1019) (0.0857) (0.1193) (0.1197) (0.0831) (0.0746)Median Family Income $65,134 $61,698 $65,877 $73,339 $56,169 $73,257 $65,147 $54,257 $62,403 $60,344 $74,197

(32154) (31425) (26990) (34312) (25680) (32519) (31973) (32133) (41752) (18432) (33824)% College Educated 0.3650 0.3298 0.3872 0.4283 0.2867 0.4241 0.3543 0.2849 0.3312 0.3450 0.4252

(0.1943) (0.1876) (0.1904) (0.2112) (0.1693) (0.1896) (0.1862) (0.1764) (0.2212) (0.1623) (0.1731)

Table 4: Unit and Neighborhood Characteristics for Rental Properties in our Sample

(0.1943) (0.1876) (0.1904) (0.2112) (0.1693) (0.1896) (0.1862) (0.1764) (0.2212) (0.1623) (0.1731)Notes: Standard deviations are shown in parenthesis. Unit characteristics are recorded from advertisements made on Craigslist.org. Neighborhood characteristics are from the 2000 census for units where we are able to match the address (or map link) to a census tract. Median rents are from fiscal year 2010 estimates by the Department of Housing and Urban Development (HUD). We match a unique median rent to each unit by city and number of bedrooms of the unit to determine if the unit rents for above or below the city median.

(1) (2) (3) (4) (5) (6) (7) (8)White African American (1) ‐ (2) Respond to Neither Respond to Both White Only African American Only (6) ‐ (7)

All Audits 57.12% 50.81% 6.30% 36.50% 43.04% 12.50% 7.96% 4.54%[4729] [4727] p=0.0000 [1151] [1357] [394] [251] p=0.0000

Atlanta 61.71% 56.75% 4.96% 29.76% 51.19% 9.82% 9.23% 0.60%[504] [504] p=0.1096 [100] [172] [33] [31] p=0.7927

Boston 64.09% 51.98% 12.10% 33.93% 44.35% 14.88% 6.85% 8.04%[504] [504] p=0.0000 [114] [149] [50] [23] p=0.0008

Chicago 51.89% 46.92% 4.97% 39.70% 39.70% 12.54% 8.06% 4.48%[503] [503] p=0.1150 [133] [133] [42] [27] p=0.0566

Dallas 51.85% 51.27% 0.59% 39.09% 34.55% 13.64% 12.73% 0.91%[162] [158] p=0.9165 [43] [38] [15] [14] p=0.8420

Washington D.C. 64.88% 58.93% 5.95% 27.98% 49.40% 14.29% 8.33% 5.95%[504] [504] p=0.0518 [94] [166] [48] [28] p=0.0149

Houston 43.20% 38.59% 4.61% 53.09% 31.44% 10.82% 4.64% 6.19%[294] [298] p=0.2543 [103] [61] [21] [9] p=0.0226

Los Angeles 58.62% 47.25% 11.37% 36.47% 42.86% 14.59% 6.08% 8.51%[493] [491] p=0.0004 [120] [141] [48] [20] p=0.0003

New York 49.47% 43.65% 5.82% 42.26% 35.71% 13.10% 8.93% 4.17%[756] [756] p 0 0233 [213] [180] [66] [45] p 0 0346

Table 5: Response Rate and Landlord Level Response by Race of Home‐Seeker, Across Cities

Overall Response Rate Response at Landlord Level

[756] [756] p=0.0233 [213] [180] [66] [45] p=0.0346Seattle 62.30% 58.13% 4.17% 32.74% 47.92% 10.42% 8.93% 1.49%

[504] [504] p=0.1767 [110] [161] [35] [30] p=0.5141San Francisco 57.54% 53.37% 4.17% 36.01% 46.13% 10.71% 7.14% 3.57%

[504] [504] p=0.1835 [121] [155] [36] [24] p=0.1045Notes: The number of observations, denoted with [] in columns (1) and (2) reflect the number of e‐mails sent. The number of observations in columns (4)‐(7) reflect the number of landlords that respond to an inquiry from neither, both, or one of the racial groups. The denominator for the percentages in columns (1) and (2) are e‐mails sent by each respective racial group. The denominator for the percentages in columns (4)‐(7) are the total number of African American/white audits.

(1) (2) (3) (4) (5) (6) (7) (8)Group 1 Group 2 (1) ‐ (2) Respond to Neither Respond to Both Group 1 Only Group 2 Only (6) ‐ (7)

White, High / African American, High 56.63% 52.62% 4.01% 37.97% 41.27% 11.27% 9.49% 1.77%[2368] [2364] p‐value=0.0058 [300] [326] [89] [75] p=0.2485

White, High / White, Low 56.63% 57.60% ‐0.97% 31.22% 51.78% 9.01% 7.99% 1.02%[2368] [2361] p‐value=0.5029 [246] [408] [71] [63] p=0.4700

White, High / African American, Low 56.63% 49.01% 7.62% 36.42% 40.86% 15.74% 6.98% 8.76%[2368] [2363] p‐value=0.0001 [287] [322] [124] [55] p=0.0001

White, Low / African American, High 57.60% 52.62% 4.98% 36.34% 47.01% 9.15% 7.50% 1.65%[2361] [2364] p‐value=0.0006 [266] [370] [72] [59] p=0.2355

African American, High / African American, Low 52.62% 49.01% 3.62% 40.66% 38.12% 11.05% 6.73% 4.32%[2364] [2363] p‐value=0.0129 [320] [300] [87] [53] p=0.0026

African American, Low / White, Low 49.01% 57.60% ‐8.60% 35.28% 43.02% 7.87% 13.83% ‐5.96%[2363] [2361] p‐value=0.0001 [278] [339] [62] [109] p=0.0001

Table 6: Response Rate and Landlord Level Response by Race and Class of Home‐Seeker

Notes: The number of observations, denoted with [] in columns (1) and (2) reflect the number of e‐mails sent. The number of observations in columns (4)‐(7) reflect the number of landlords that respond to an inquiry from neither, both, or one of the racial groups. The denominator for the percentages in columns (1) and (2) are e‐mails sent by each respective race‐class group. The denominator for the percentages in columns (4)‐(7) are the total number of audits for that combination.


(1) (2) (3) (4) (5) (6) (7) (8) (9)Race= 1 if African American ‐0.0630 ‐0.0559 ‐0.222 ‐0.0401 ‐0.0190 ‐0.119 ‐0.0364 0.0271 ‐0.256

[0.000] [ 0.079] [0.000] [0.002] [0.670] [0.003] [ 0.077] [0.709] [0.001]Class= 1 if Low 0.0098 0.0365 ‐0.0585 ‐0.0265 ‐0.0251 ‐0.185

[.01293] [ 0.401] [0.160] [0.200] [0.725] [0.013]Order= 1 if Second ‐0.0851 ‐0.0834 ‐0.47

[0.000] [ 0.228] [0.000]Race*Class ‐0.0459 ‐0.1851 ‐0.188 ‐0.0328 ‐0.0911 ‐0.122

[ 0.012] [ 0.237] [0.001] [0.262] [ 0.369] [0.241]Race*Order ‐0.0077 ‐0.0707 0.208

[0.808] [0.525] [0.086]Class*Order 0.0715 0.1200 0.157

[ 0.024] [0.240] [0.179]Race*Class*Order ‐0.0258 0.0168 ‐0.0571

[0.569] [ 0.913] [0.739]N 9456 830 1838 9456 830 1838 9456 830 1838Pseudo R2 0.0029 0.0570 0.0251 0.0034 0.0579 0.0400 0.0067 0.0633 0.1145Notes: All coefficients are expressed in terms of marginal effects. Columns (1), (3), (4), (6), (7) and (9) show estimates without control variables, of these columns (3), (6), and (9) show estimates with landlord fixed effects . Columns (2), (5), and (8) control for the city, number of bedrooms and bathrooms, square footage of the unit, an indicator if the unit is an apartment or single family home, monthly rent, an indicator if the monthly rent is above the median, percentage of residents living below the poverty line, percentage of residents with at least a college degree, percentage of white residents, and an indicator if the neighborhood is a "tipping" neighborhood (between 80 and 95 percent white). Standard errors in all Probits are clustered at the landlord level, and p‐values are shown in [].

Table 7: Marginal Effects on the Likelihood of Response by Race, Class, and Order


Less than .50 56.98% 48.58% 8.41% 39.07% 38.84% 13.95% 8.14% 5.81%[630] [632] p=0.0028 [430] [168] [167] [60] p=0.0065

.50 to .80 61.29% 56.46% 4.83% 50.29% 32.43% 10.86% 6.43% 4.43%[1072] [1068] p=0.0233 [700] [352] [227] [76] p=0.0032

.80 to .95 61.87% 51.25% 10.62% 44.77% 33.42% 14.41% 7.40% 7.02%[1154] [1202] p=0.0001 [784] [351] [262] [113] p=0.0001

.95 to 1 51.84% 48.00% 3.84% 39.23% 39.95% 11.70% 9.12% 2.58%[1873] [1825] p=0.0195 [1239] [486] [495] [145] p=0.0353

Table 8: Response Rate and Landlord Level Response by Absolute Percentage of White Residents in Surrounding Neighborhood


Notes: The number of observations, denoted with [] in columns (1) and (2) reflect the number of e‐mails sent. The number of observations in columns (4)‐(7) reflect the number of landlords that respond to an inquiry from neither, both, or one of the racial groups. The denominator for the percentages in columns (1) and (2) are e‐mails sent by each respective race. The denominator for the percentages in columns (4)‐(7) is the total number of African American/white audits. The percentage of white residents in a census tract comes from the 2000 census.


Single Family 59.67% 56.44% 3.23% 47.54% 32.65% 10.14% 9.67% 0.48%[977] [939] p=0.1521 [300] [206] [64] [61] p=0.7774

Duplex 56.52% 51.61% 4.91% 42.11% 36.84% 10.53% 10.53% 0.00%[115] [155] p=0.4238 [40] [35] [10] [10] p=1.0000

Townhouse 63.52% 61.67% 1.86% 51.88% 29.38% 8.75% 10.00% ‐1.25%[244] [240] p=0.6728 [83] [47] [14] [16] p=0.7013

Condominiums 65.56% 56.47% 9.09% 50.50% 27.72% 16.83% 4.95% 11.88%[453] [425] p=0.0057 [153] [84] [51] [15] p=0.0001

Apartment 57.26% 47.58% 9.68% 40.56% 38.50% 13.52% 7.43% 6.09%[1647] [1673] p=0.0001 [453] [430] [151] [83] p=0.0001

Shared Room 64.47% 52.24% 12.23% 47.96% 31.63% 15.31% 5.10% 10.20%[152] [134] p=0.0360 [47] [31] [15] [5] p=0.0183


Notes: The number of observations, denoted with [] in columns (1) and (2) reflect the number of e‐mails sent. The number of observations in columns (4) ‐ (7) reflect the number of landlords that respond to an inquiry from neither, both, or one of the racial groups. The denominator for the percentages in columns (1) and (2) are e‐mails sent by each respective race. The denominator for the percentages in columns (4) ‐ (7) are the total number of African American/white audits. The type of rental unit is identified by the authors from the description in the unit advertisement.

Table 9: Response Rate and Landlord Level Response by Type of Rental Unit

Unit Characteristics N Marginal Effect on Response for African American Names p‐valuePercentage of Median Rent 7586 ‐0.0014 [0.378]Square Footage (in thousands) 1454 0.0574 [0.024]Discount 7768 0.0303 [0.293]Address Included 7810 ‐0.0225 [0.371]# of Bedrooms 7650 0.0194 [0.037]# of Bathrooms 5074 0.0204 [0.218]Apartment 7024 ‐0.0478 [0.027]Single Family Home 7032 0.0513 [0.031]Shared Room 7502 ‐0.0602 [0.288]Neighborhood CharacteristicsPercent White 5908 0.0035 [0.938]Tipping Neighborhood 5908 ‐0.0525 [0.028]Percent College Educated 5908 0.0572 [0.334]Median Income (in thousands) 5908 0.0003 [0.366]Percent Below Poverty 5908 ‐0.1540 [0.177]Employment Rate 5908 0.1065 [0.318]Percent Migrants 5908 0.1103 [0.178]Notes: Probit regressions are at the e‐mail level and cluster all standard errors at the landlord level. The marginal effect of all characteristics is calculated from a probit regression of the respond variable on a dummy for an African American sounding name, the characteristic, and the interaction between the characteristic and the African American sounding name dummy. We report the marginal effect of the interaction term. The percent of migrants in a neighborhood is measured using the number of residents who lived elsewhere in 1995, divided by total residents in 2000.

Table 10: Marginal Effects of Unit and Neighborhood Characteristics on Response to e‐mail Inquiries from African American Sounding Names

(1) (2) (3) (4) (5)Neither Both White Only African American Only (3) ‐ (4)

All Audits 19.68% 68.98% 6.12% 5.16% 0.96%[267] [936] [83] [70] p=0.2793

Atlanta 19.19% 66.28% 7.56% 6.40% 1.16%[33] [114] [13] [11] p=0.6721

Boston 10.07% 77.85% 7.38% 4.70% 2.68%[15] [116] [11] [7] p=0.3307

Chicago 30.08% 55.64% 9.02% 5.26% 3.76%[40] [74] [12] [7] p=0.2339

Dallas 13.16% 71.05% 5.26% 10.53% ‐5.26%[5] [27] [2] [4] p=0.3949

Washington D.C. 15.06% 75.90% 5.42% 3.61% 1.81%[25] [126] [9] [6] p=0.4279

Houston 22.95% 70.49% 3.28% 3.28% 0.00%[14] [43] [2] [2] p=1.0000

Los Angeles 22.70% 63.12% 8.51% 5.67% 2.84%[32] [89] [12] [8] p=0.3534

New York 31.67% 55.00% 5.56% 7.78% ‐2.22%[57] [99] [10] [14] p=0.3980

Seattle 14.29% 78.88% 2.48% 4.35% ‐1.86%[23] [127] [4] [7] p=0.3574

San Francisco 14.84% 77.42% 5.16% 2.58% 2.58%[23] [120] [8] [4] p=0.2389

Table 11: Landlord Level: Describes Unit as Available by Race of Home‐Seeker

Notes: The number of observations reflect the number of landlords that describe the rental unit as available to an inquiry from neither, both, or one of the racial groups. The denominator is landlords who are sent inquires from both African American and white potential tenants and reply to both.

(1) (2) (3) (4) (5)Neither Both White Only African‐American Only (3) ‐ (4)

All Audits 27.81% 58.07% 7.27% 6.84% 0.43%[260] [543] [68] [64] p=0.7180

Atlanta 39.47% 46.49% 7.89% 6.14% 1.75%[45] [53] [9] [7] p=0.6041

Boston 25.86% 57.76% 9.48% 6.90% 2.59%[30] [67] [11] [8] p=0.4726

Chicago 27.03% 54.05% 8.11% 10.81% ‐2.70%[20] [40] [6] [8] p=0.5743

Dallas 33.33% 44.44% 11.11% 11.11% 0.00%[9] [12] [3] [3] p=1.0000

Washington D.C. 24.60% 59.52% 8.73% 7.14% 1.59%[31] [75] [11] [9] p=0.6411

Houston 46.51% 37.21% 9.30% 6.98% 2.33%[20] [16] [4] [3] p=0.6933

Los Angeles 25.84% 61.80% 8.99% 3.37% 5.62%[23] [55] [8] [3] p=0.1196

New York 23.23% 65.66% 7.07% 4.04% 3.03%[23] [65] [7] [4] p=0.3520

Seattle 28.35% 58.27% 3.15% 10.24% ‐7.09%[36] [74] [4] [13] p=0.0238

San Francisco 19.17% 71.67% 4.17% 5.00% ‐0.83%[23] [86] [5] [6] 0 7576

Table 12: Landlord Level: Invites Home‐Seeker For Viewing by Race of Home‐Seeker

[23] [86] [5] [6] p=0.7576Notes: The number of observations reflect the number of landlords that describe the rental unit as available to an inquiry from neither, both, or one of the racial groups. The denominator is the landlords who are sent e‐mail inquires from both African American and white potential tenants, reply to both, and describe the unit as available in their response.

(1) (2) (3) (4) (5) (6) (7) (8)White African‐American (1) ‐ (2) Respond to Neither Respond to Both White Only African‐American Only Ho: Rw‐RAA=0

57.14% 49.96% 7.18% 37.01% 42.24% 13.36% 7.40% 5.95%

[4727] [2644] p= 0.0000 [665] [759] [240] [133] p= 0.0000

57.00% 50.82% 6.18% 36.80% 43.42% 12.13% 7.65% 4.48%

[4200] [3656] p= 0.0000 [789] [931] [260] [164] p= 0.0000

56.27% 50.42% 5.85% 38.71% 42.49% 10.90% 7.90% 3.00%

[3172] [2608] p= 0.0000 [451] [495] [127] [92] p= 0.0130

58.83% 51.30% 7.53% 35.21% 43.36% 13.43% 8.00% 5.43%

[1557] [2119] p= 0.0000 [700] [862] [267] [159] p= 0.0000

57.70% 50.81% 6.88% 36.09% 43.38% 12.42% 8.11% 4.31%

[3489] [4727] p= 0.0000 [837] [1006] [288] [188] p= 0.0000

56.01% 50.83% 5.18% 38.11% 42.63% 11.81% 7.45% 4.36%

[1948] [4295] p=0.0001 [455] [509] [141] [89] p=0.0003

Excluding four most unique and four most common names (Matthew, Todd, Brett, Brad, Hakim, Rasheed, Tremayne, Kareem)

Audits that only contain either a top four most common or most unique name

Notes: The number of observations, denoted with [] in columns (1) and (2) reflect the number of e‐mails sent. The number of observations in columns (4)‐(7) reflect the number of landlords that respond to an inquiry from neither, both, or one of the racial groups. The denominator for the percentages in columns (1) and (2) are e‐mails sent by each respective racial group. The denominator for the percentages in columns (4)‐(7) are the total number of African American/white audits. Name popularity or uniqueness is according to rankings of the most popular names for babies born between 1974 and 1979 inclusive (the years of the Bertrand and Mullainathan birth certificate data) and comes from the Social Security Administration website at: http://www.ssa.gov/OACT/babynames/

Table 13: Results Excluding Religious or Unique Names

Overall Response Rate Response at landlord level

Excluding Muslim Sounding Names (Hakim, Jamal, Kareem, Rasheed)

Excluding two most unique and two most common names (Matthew, Todd, Rasheed, Hakim)

Excluding names ranked in the top 100 most common by State

Excluding names ranked in the top 50 most common by State

Name National Georgia Massachusetts Illinois Texas D.C California New York WashingtonMatthew 8 12 5 7 14 11 9 11 6Todd 50 89 42 43 NA 74 73 48 50Brett 103 NA 78 82 NA NA NA 88 83Brad 109 65 81 51 66 91 100 88 62Jay 122 NA 86 NA NA NA NA NA 89Tyrone 155 89 NA NA NA 59 NA 94 NANeil 169 NA 84 NA NA NA NA NA NAJermaine 174 88 NA NA NA 67 NA 91 NAGeoffrey 182 NA 93 NA NA NA NA NA 99Greg 252 31 34 34 47 23 38 29 46Brendan 258 NA 68 NA NA 94 NA 99 NALeroy 264 NA NA NA NA NA NA NA NAJamal 336 NA NA NA NA 68 NA NA NADarnell 349 NA NA NA NA 86 NA NA NAKareem 403 NA NA NA NA NA NA NA NATremayne 669 NA NA NA NA NA NA NA NARasheed 767 NA NA NA NA NA NA NA NAHakim 882 NA NA NA NA NA NA NA NA

Table 14: First Name Popularity at National and States in Sample Level

Name popularity rankings are average rankings for babies born between 1974 and 1979 inclusive. Popularity rankings for the nation reflect an appearance in the top 1,000 most common names, popularity rankings by state reflect an appearance in the top 100 most common names. All popularity rankings are specific to males and are taken from the Social Security Administration at http://www.ssa.gov/OACT/babynames/

Name ATL BOS CHI DAL DC HOU LA NYC SEA SFBrad Davis [0.7680] [0.4371] [0.3929] [0.5454] [0.2828] [0.8050] [0.8897] [0.4258] [0.8229] [0.6294]

Brendan Ryan [0.7292] [0.1093] [0.6879] [0.3797] [0.2130] [0.5360] [0.6915] [0.4047] [0.8313] [0.4655]Matthew O’Brien [0.6854] [0.6933] [0.6006] [0.2157] [0.3999] [0.7506] [0.1393] [0.6787] [0.7189] [0.5645]

Todd Jones [0.6741] [0.4623] [0.3601] [0.8814] [0.3748] [0.8217] [0.1253] [0.8289] [0.4389] [0.9902]Brett Murphy [0.6547] [0.9472] [0.9608] +[0.0169] [0.6629] [0.8866] [0.4998] [0.5548] [0.2429] [0.5403]Neil Baker [0.2763] [0.5366] [0.4403] [0.6177] [0.8395] [0.5446] ‐[0.0820] ‐[0.0650] [0.4096] [0.6087]

Geoffrey McCarthy [0.2956] [0.4957] [0.6209] [0.3172] [0.3557] [0.6902] [0.5740] [0.2700] [0.6496] [0.9457]Greg Young [0.5122] [0.6689] [0.7642] [0.5363] [0.4757] [0.3858] [0.7606] [0.7649] [0.7760] [0.9897]Jay Wright [0.4102] [0.2809] [0.7997] [0.7903] [0.8383] [0.2252] [0.1475] [0.7591] [0.5793] [0.7054]

Jermaine Jackson [0.1022] [0.1046] [0.1081] [0.2567] [0.2140] [0.6245] [0.5218] [0.2017] [0.9765] ‐[0.0383]Kareem Hall [0.7352] [0.1751] [0.9486] [0.8955] +[0.0865] [0.9896] [0.3568 [0.5095] [0.1463] [0.8641]Leroy Parker +[0.0563] [0.1309] [0.2823] [0.7615] [0.5129] [0.3412] [0.8024] [0.2145] [0.4218] [0.9399]

Hakim Washington [0.4377] [0.6892] [0.6562] [0.9234] [0.9099] [0.9040] [0.4944] [0.4278] [0.4255] [0.2135]Jamal Robinson [0.5187] [0.6615] [0.5736] [0.7416] [0.5824] [0.6117] +[0.0544] [0.1142] [0.8607] [0.7280]Darnell Johnson [0.6434] [0.5919] [0.5644] [0.6366] [0.6058] [0.2382] [0.4605] [0.7306] [0.5989] [0.1550]Rasheed Jones [0.5528] [0.7307] [0.4101] +[0.0312] [0.4749] [0.2428] [0.1406] [0.2057] [0.7993] [0.7932]

Tremayne Williams [0.6195] [0.5581] [0.2386] [0.1553] [0.2428] [0.7195] [0.4450] [0.8198] [0.7982] [0.7527]Tyrone Cooper [0.8644] ‐[0.0551] [0.4747] [0.8333] [0.5374] ‐[0.0499] [0.7009] [0.1429] [0.4939] [0.6133]

Table 15: P‐value of Tests for Difference in Response Rate Between Specific Name and Other Names of Same Race Within City

Notes: Table reports p‐values for a test of symmetry between the response rate from e‐mail inquires made from a specific name in a specific city and all other names in that race/city grouping. P‐values shown in bold are statistically significant at the 5 percent level, those shown in italics are statistically significant at the 10 percent level. A (+) next to a p‐value means that particular name has a more favorable response rate than other names representing the same race in that particular city. A (‐) next to a p‐value means that particular name has a less favorable response rate than other names representing the same race in that particular city.


All Audits 57.01% 50.83% 6.18% 36.54% 42.94% 12.50% 8.02% 4.48%[4708] [4474] p=0.0000 [1084] [1274] [371] [238] p=0.0000

Atlanta 61.71% 55.19% 6.52% 30.74% 49.66% 10.81% 8.78% 2.03%[504] [453] p=0.0409 [91] [147] [32] 26] p=0.4068

Boston 64.09% 53.51% 10.58% 33.11% 45.03% 14.57% 7.28% 7.28%[504] [456] p=0.0009 [100] [136] [44] [22] p=0.0041

Chicago 51.89% 46.92% 4.97% 39.70% 39.70% 12.54% 8.06% 4.48%[503] [503] p=0.1149 [133] [133] [42] [27] p=0.0566

Dallas 47.89% 47.45% 0.44% 44.44% 33.33% 12.22% 10.00% 2.22%[142] [137] p=0.9411 [40] [30] [11] [9] p=0.6353

Washington D.C. 64.88% 58.93% 5.95% 27.98% 49.40% 14.29% 8.33% 5.95%[504] [504] p=0.0517 [94] [166] [48] [28] p=0.0148

Houston 43.20% 40.44% 2.76% 52.00% 33.14% 9.71% 5.14% 4.57%[294] [272] p=0.5067 [91] [58] [17] [9] p=0.1030

Los Angeles 58.62% 45.74% 12.88% 37.12% 41.14% 15.38% 6.35% 9.03%[493] [446] p=0.0001 [111] [123] [46] [19] p=0.0004

New York 49.47% 43.65% 5.82% 42.26% 35.71% 13.10% 8.93% 4.17%[756] [756] p=0.0233 [213] [180] [66] [45] p=0.0346

Seattle 62.30% 58.13% 4.17% 32.74% 47.92% 10.42% 8.93% 1.49%[504] [504] p=0.1766 [110] [161] [35] [30] p=0.5141

San Francisco 57.54% 55.30% 2.23% 34.35% 47.62% 10.20% 7.82% 2.38%[504] [443] p=0.4888 [101] [140] [30] [23] p=0.3134

Table 16: Response Rate and Landlord Level Response by Race of Home‐Seeker, Excluding Names with Unusual Performance (significant at 5 percent level)


Notes: The number of observations, denoted with [] in columns (1) and (2) reflect the number of e‐mails sent. The number of observations in columns (4)‐(7) reflect the number of landlords that respond to an inquiry from neither, both, or one of the racial groups. The denominator for the percentages in columns (1) and (2) are e‐mails sent by each respective racial group. The denominator for the percentages in columns (4)‐(7) is the total number of African‐American/white audits. These results exclude e‐mails sent and audits containing at least one e‐mail from the following name/city combinations: Leroy Parker in Atlanta, Tyrone Cooper in Boston, Brett Murphy in Dallas, Rasheed Jones in Dallas, Tyrone Cooper in Houston, Jamal Robinson in Los Angeles, and Jermaine Jackson in San Francisco.


All Audits 54.63% 48.25% 6.38% 39.23% 40.51% 12.20% 8.05% 4.15%[3778] [3780] p=0.0000 [984] [1016] [306] [202] p=0.0000

Atlanta 62.50% 54.68% 7.82% 30.74% 48.15% 11.85% 9.26% 2.59%[400] [406] p=0.0242 [83] [130] [32] [25] p=0.3269

Boston 61.92% 50.12% 11.80% 36.30% 41.64% 14.59% 7.47% 7.12%[407] [419] p=0.0006 [102] [117] [41] [21] p=0.0071

Chicago 49.11% 43.00% 6.11% 43.97% 36.19% 12.84% 7.00% 5.84%[395] [393] p=0.0853 [113] [93] [33] [18] p=0.0269

Dallas 50.00% 49.28% 0.72% 40.43% 30.85% 15.96% 12.77% 3.19%[134] [138] p=0.9049 [38] [29] [15] [12] p=0.5327

Washington D.C. 63.50% 58.76% 4.74% 28.57% 49.60% 12.70% 9.13% 3.57%[400] [388] p=0.1725 [72] [125] [32] [23] p=0.1985

Houston 41.92% 32.74% 9.19% 54.30% 29.80% 11.26% 4.64% 6.62%[229] [223] p=0.0436 [82] [45] [17] [7] p=0.0334

Los Angeles 53.19% 44.82% 8.37% 41.53% 38.31% 13.31% 6.85% 6.45%[376] [386] p=0.0208 [103] [95] [33] [17] p=0.0170

New York 46.28% 40.94% 5.35% 44.93% 33.18% 12.67% 9.22% 3.46%[646] [640] p=0.0532 [195] [144] [55] [40] p=0.1029

Seattle 59.38% 55.88% 3.49% 36.59% 46.75% 8.94% 7.72% 1.22%[384] [374] p=0.3306 [90] [115] [22] [19] p=0.6246

San Francisco 55.04% 50.85% 4.19% 38.55% 44.73% 9.45% 7.27% 2.18%[407] [413] p=0.2295 [106] [123] [26] [20] p=0.3554

Table 17: Response Rate and Landlord Level Response by Race of Home‐Seeker, Using only anonymous e‐mail addressesOverall Response Rate Response at Landlord Level

Notes: The number of observations, denoted with [] in columns (1) and (2) reflect the number of e‐mails sent. The number of observations in columns (4)‐(7) reflect the number of landlords that respond to an inquiry from neither, both, or one of the racial groups. The denominator for the percentages in columns (1) and(2) are e‐mails sent by each respective racial group. The denominator for the percentages in columns (4)‐(7) is the total number of African‐American/white audits. These results include only e‐mails sent to landlords using anonymous craigslist e‐mail addresses of the form, hous‐aaaa‐##########@craigslist.org, where aaaa is a four character letter code and # is the posting id of the advertisement.

Craigslist Sample AHS/Census* Statistically DifferentTotal Housing Units 4728 2835 actual, 6,724,157 weightedBedrooms 2.04 1.78 Yes

( 1.1195) (0.9435)Bathrooms 1.61 1.23 Yes

(.7699) (0.4963)Single Family Homes 0.2724 0.1321 Yes

(.4453) (0.3387)Townhouse 0.0689 0.0438 No

(.2533) (0.2048)Condo 0.1250 0.0898 No

(.3308) (0.2859)Apartment 0.4725 0.7397 Yes

(.4993) (0.4389)Monthly Rent $1,492 $1,025 Yes

(811) (666)Greater than Median Area Rent 0.4594 0.2608 Yes

(.4984) (0.4391)Square Footage 1361 1149 Yes

(1209) (2020)% African American Residents 0.1438 0.1647 No

(0.2286) (0.2632)% White Residents 0.6847 0.6267 Yes

(0.2535) (0.2904)% Residents Under Poverty Line 0.1306 0.1237 No

(0.1045) (0.1175)

Table 18: Unit and Neighborhood Characteristics for Rental Properties in our Sample Versus the Population of Rental Units

( ) ( )Median Family Income $65,134 $60,881 Yes

(32154) (30404)% College Educated 0.3650 0.2624 Yes

(0.1943) (0.1761)

The column statistically different is based on a difference in means t test between the means of the craigslist and AHS/Census samples

*AHS Sample includes only units identified as part of the Metropolitan Statistical Areas from our Craigslist sample, Census sample includes all census tracts identified as part of these same areas.The AHS does not seperately identify homes as duplexes or shared rooms, so these categories are omitted

05

1015

200

510

1520

ATLBOS

CHIDAL

DCHOU

LA***NYC

SEASF

ATLBOS

CHIDAL

DCHOU

LA*NYC

SEASF

ATLBOS***

CHI**DAL

DC***HOU

LANYC**

SEA**SF

ATLBOS

CHIDAL

DCHOU**

LA**NYC

SEA*SF

ATL*BOS

CHIDAL

DCHOU

LANYC**

SEASF

ATLBOS**

CHIDAL

DC**HOU

LA**NYC

SEASF

(1) White High--African American High (2) White High--White Low (3) White High--African American Low

(4) African American High--White Low (5) African American High--African American Low (6) African American Low--White Low

Group 1 Only Group 2 Only

Figure 1: Landlord Level Response by Race and Class Across Cities

Note: * denotes significant difference at 10% level, ** at 5% level and *** at 1% level

Do Landlords Discriminate in the Rental Housing Market...

Documents

Transcript of Do Landlords Discriminate in the Rental Housing Market...